app.py

import json

import gradio as gr
from huggingface_hub import snapshot_download
from omegaconf import OmegaConf
from vosk import KaldiRecognizer, Model


def load_vosk(model_id: str):
    model_dir = snapshot_download(model_id)
    return Model(model_path=model_dir)


OmegaConf.register_new_resolver("load_vosk", load_vosk)

models_config = OmegaConf.to_object(OmegaConf.load("configs/models.yaml"))


def automatic_speech_recognition(
    model_id: str, dialect_id: str, stream: str, new_chunk: str
):
    if isinstance(models_config[model_id]["model"], dict):
        model = models_config[model_id]["model"][dialect_id]
    else:
        model = models_config[model_id]["model"]

    sample_rate, audio_array = new_chunk
    if audio_array.ndim == 2:
        audio_array = audio_array[:, 0]

    audio_bytes = audio_array.tobytes()

    if stream is None:
        rec = KaldiRecognizer(model, sample_rate)
        rec.SetWords(True)
        result = []
    else:
        rec, result = stream

    if rec.AcceptWaveform(audio_bytes):
        text_result = json.loads(rec.Result())["text"]
        if text_result != "":
            result.append(text_result)
        partial_result = ""
    else:
        partial_result = json.loads(rec.PartialResult())["partial"] + " "

    if len(result) > 0:
        output_text = "，".join(result) + "，" + partial_result
    else:
        output_text = partial_result

    return (rec, result), output_text


def when_model_selected(model_id: str):
    model_config = models_config[model_id]

    if "dialect_mapping" not in model_config:
        return gr.update(visible=False)

    dialect_drop_down_choices = [
        (k, v) for k, v in model_config["dialect_mapping"].items()
    ]

    return gr.update(
        choices=dialect_drop_down_choices,
        value=dialect_drop_down_choices[0][1],
        visible=True,
    )


demo = gr.Blocks(
    title="臺灣客語語音辨識系統",
    css="@import url(https://tauhu.tw/tauhu-oo.css);",
    theme=gr.themes.Default(
        font=(
            "tauhu-oo",
            gr.themes.GoogleFont("Source Sans Pro"),
            "ui-sans-serif",
            "system-ui",
            "sans-serif",
        )
    ),
)

with demo:
    default_model_id = list(models_config.keys())[0]
    model_drop_down = gr.Dropdown(
        models_config.keys(),
        value=default_model_id,
        label="模型",
    )

    dialect_drop_down = gr.Dropdown(
        choices=[
            (k, v)
            for k, v in models_config[default_model_id]["dialect_mapping"].items()
        ],
        value=list(models_config[default_model_id]["dialect_mapping"].values())[0],
        label="腔調",
    )

    model_drop_down.input(
        when_model_selected,
        inputs=[model_drop_down],
        outputs=[dialect_drop_down],
    )

    gr.Markdown(
        """
        # 臺灣客語語音辨識系統
        ### Taiwanese Hakka Automatic-Speech-Recognition System
        ### 研發
        - **[李鴻欣 Hung-Shin Lee](mailto:[email protected])（[聯和科創](https://www.104.com.tw/company/1a2x6bmu75)）**
        - **[陳力瑋 Li-Wei Chen](mailto:[email protected])（[聯和科創](https://www.104.com.tw/company/1a2x6bmu75)）**
        ### 合作單位
        - **[國立聯合大學智慧客家實驗室](https://www.gohakka.org)**
        """
    )
    state = gr.State()
    audio = gr.Audio(
        label="錄音",
        type="numpy",
        format="wav",
        waveform_options=gr.WaveformOptions(
            sample_rate=16000,
        ),
        sources=["microphone"],
        streaming=True,
    )
    gr.Interface(
        automatic_speech_recognition,
        inputs=[
            model_drop_down,
            dialect_drop_down,
            state,
            audio,
        ],
        outputs=[
            state,
            gr.Text(interactive=False, label="客語漢字"),
        ],
        live=True,
        stream_every=0.25,
        clear_btn=None,
        # flagging_mode="auto",
    )


demo.launch()